Building Footprint Extraction for the USA¶

Install package¶

To use the geoai-py package, ensure it is installed in your environment. Uncomment the command below if needed.

In [1]:

# %pip install geoai-py

# %pip install geoai-py

Import libraries¶

In [2]:

import geoai

import geoai

Download sample data¶

In [3]:

raster_url = (
    "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train.tif"
)
vector_url = "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train_buildings.geojson"

raster_url = ( "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train.tif" ) vector_url = "https://huggingface.co/datasets/giswqs/geospatial/resolve/main/naip_train_buildings.geojson"

In [4]:

raster_path = geoai.download_file(raster_url)

raster_path = geoai.download_file(raster_url)

naip_train.tif:   0%|          | 0.00/12.1M [00:00<?, ?B/s]

naip_train.tif:  93%|█████████▎| 11.2M/12.1M [00:00<00:00, 118MB/s]

naip_train.tif: 100%|██████████| 12.1M/12.1M [00:00<00:00, 117MB/s]

In [5]:

vector_path = geoai.download_file(vector_url)

vector_path = geoai.download_file(vector_url)

naip_train_buildings.geojson:   0%|          | 0.00/456k [00:00<?, ?B/s]

naip_train_buildings.geojson: 100%|██████████| 456k/456k [00:00<00:00, 22.7MB/s]

Initialize building footprint extraction pretrained model¶

The pretained model is adapted from the Esri building footprint extraction model for the USA. Credits to Esri for the model.

In [6]:

extractor = geoai.BuildingFootprintExtractor()

extractor = geoai.BuildingFootprintExtractor()

Model path not specified, downloading from Hugging Face...

Model downloaded to: /home/runner/.cache/huggingface/hub/models--giswqs--geoai/snapshots/c4b9e7e7686c41a4fa7e4d5dd97f03853e088e4d/building_footprints_usa.pth
Model loaded successfully

Extract building footprints¶

Option 1: Extract building footprints as raster¶

In [7]:

mask_path = extractor.save_masks_as_geotiff(
    raster_path=raster_path,
    output_path="building_masks.tif",
    confidence_threshold=0.5,
    mask_threshold=0.5,
)

mask_path = extractor.save_masks_as_geotiff( raster_path=raster_path, output_path="building_masks.tif", confidence_threshold=0.5, mask_threshold=0.5, )

Processing masks with parameters:
- Confidence threshold: 0.5
- Chip size: (512, 512)
- Mask threshold: 0.5
Dataset initialized with 3 rows and 7 columns of chips
Image dimensions: 2503 x 1126 pixels
Chip size: 512 x 512 pixels
Overlap: 25.0% (stride_x=384, stride_y=384)
CRS: EPSG:26911

Processing raster with 6 batches

  0%|          | 0/6 [00:00<?, ?it/s]

 17%|█▋        | 1/6 [00:11<00:55, 11.10s/it]

 33%|███▎      | 2/6 [00:23<00:47, 11.98s/it]

 50%|█████     | 3/6 [00:34<00:34, 11.55s/it]

 67%|██████▋   | 4/6 [00:45<00:22, 11.37s/it]

 83%|████████▎ | 5/6 [00:56<00:11, 11.21s/it]

100%|██████████| 6/6 [00:59<00:00,  8.32s/it]

100%|██████████| 6/6 [00:59<00:00,  9.91s/it]

Object masks saved to building_masks.tif

Convert raster to vector

In [8]:

gdf = extractor.masks_to_vector(
    mask_path=mask_path,
    output_path="building_masks.geojson",
    simplify_tolerance=1.0,
)

gdf = extractor.masks_to_vector( mask_path=mask_path, output_path="building_masks.geojson", simplify_tolerance=1.0, )

Converting mask to GeoJSON with parameters:
- Mask threshold: 0.5
- Min object area: 100
- Max object area: None
- Simplify tolerance: 1.0
- NMS IoU threshold: 0.5
- Regularize objects: True
- Angle threshold: 15° from 90°
- Rectangularity threshold: 70.0%
Mask dimensions: (1126, 2503)
Mask value range: 0 to 255
Found 632 potential objects

  0%|          | 0/632 [00:00<?, ?it/s]

 12%|█▏        | 77/632 [00:00<00:00, 767.93it/s]

 25%|██▌       | 158/632 [00:00<00:00, 787.31it/s]

 38%|███▊      | 237/632 [00:00<00:00, 778.65it/s]

 50%|█████     | 316/632 [00:00<00:00, 779.39it/s]

 63%|██████▎   | 397/632 [00:00<00:00, 788.98it/s]

 75%|███████▌  | 476/632 [00:00<00:00, 781.15it/s]

 88%|████████▊ | 555/632 [00:00<00:00, 777.69it/s]

100%|██████████| 632/632 [00:00<00:00, 781.64it/s]

Created 609 valid polygons

Object count after NMS filtering: 609
Regularizing 609 objects...
- Angle threshold: 15° from 90°
- Min orthogonality: 30.0% of angles
- Min rectangularity: 70.0% of bounding box area

  0%|          | 0/609 [00:00<?, ?it/s]

 21%|██        | 127/609 [00:00<00:00, 1266.34it/s]

 43%|████▎     | 259/609 [00:00<00:00, 1292.17it/s]

 66%|██████▌   | 399/609 [00:00<00:00, 1337.69it/s]

 89%|████████▉ | 544/609 [00:00<00:00, 1380.93it/s]

100%|██████████| 609/609 [00:00<00:00, 1374.50it/s]

Regularization completed:
- Total objects: 609
- Rectangular objects: 609 (100.0%)
- Other regularized objects: 0 (0.0%)
- Unmodified objects: 0 (0.0%)
Saved 609 objects to building_masks.geojson

Option 2: Extract building footprints as vector¶

In [9]:

output_path = "naip_buildings.geojson"
gdf = extractor.process_raster(
    raster_path,
    output_path="buildings.geojson",
    batch_size=4,
    confidence_threshold=0.5,
    overlap=0.25,
    nms_iou_threshold=0.5,
    min_object_area=100,
    max_object_area=None,
    mask_threshold=0.5,
    simplify_tolerance=1.0,
)

output_path = "naip_buildings.geojson" gdf = extractor.process_raster( raster_path, output_path="buildings.geojson", batch_size=4, confidence_threshold=0.5, overlap=0.25, nms_iou_threshold=0.5, min_object_area=100, max_object_area=None, mask_threshold=0.5, simplify_tolerance=1.0, )

Processing with parameters:
- Confidence threshold: 0.5
- Tile overlap: 0.25
- Chip size: (512, 512)
- NMS IoU threshold: 0.5
- Mask threshold: 0.5
- Min object area: 100
- Max object area: None
- Simplify tolerance: 1.0
- Filter edge objects: True
- Edge buffer size: 20 pixels
Dataset initialized with 3 rows and 7 columns of chips
Image dimensions: 2503 x 1126 pixels
Chip size: 512 x 512 pixels
Overlap: 25.0% (stride_x=384, stride_y=384)
CRS: EPSG:26911

Processing raster with 6 batches

  0%|          | 0/6 [00:00<?, ?it/s]

 17%|█▋        | 1/6 [00:10<00:52, 10.52s/it]

 33%|███▎      | 2/6 [00:21<00:44, 11.05s/it]

 50%|█████     | 3/6 [00:33<00:33, 11.19s/it]

 67%|██████▋   | 4/6 [00:44<00:22, 11.26s/it]

 83%|████████▎ | 5/6 [00:56<00:11, 11.30s/it]

100%|██████████| 6/6 [00:58<00:00,  8.41s/it]

100%|██████████| 6/6 [00:58<00:00,  9.81s/it]

Objects before filtering: 711
Objects after filtering: 711
Saved 711 objects to buildings.geojson

Regularize building footprints¶

In [10]:

gdf_regularized = extractor.regularize_buildings(
    gdf=gdf,
    min_area=100,
    angle_threshold=15,
    orthogonality_threshold=0.3,
    rectangularity_threshold=0.7,
)

gdf_regularized = extractor.regularize_buildings( gdf=gdf, min_area=100, angle_threshold=15, orthogonality_threshold=0.3, rectangularity_threshold=0.7, )

Regularizing 711 objects...
- Angle threshold: 15° from 90°
- Min orthogonality: 30.0% of angles
- Min rectangularity: 70.0% of bounding box area

  0%|          | 0/711 [00:00<?, ?it/s]

 19%|█▉        | 134/711 [00:00<00:00, 1338.37it/s]

 39%|███▉      | 279/711 [00:00<00:00, 1401.94it/s]

 60%|██████    | 429/711 [00:00<00:00, 1442.18it/s]

 82%|████████▏ | 582/711 [00:00<00:00, 1476.61it/s]

100%|██████████| 711/711 [00:00<00:00, 1482.92it/s]

Regularization completed:
- Total objects: 711
- Rectangular objects: 664 (93.4%)
- Other regularized objects: 0 (0.0%)
- Unmodified objects: 47 (6.6%)

Visualize building footprints¶

In [11]:

gdf.head()

gdf.head()

Out[11]:

	geometry	confidence	class
991	POLYGON ((455092.8 5277652.2, 455092.2 5277651...	0.988162	1
957	POLYGON ((455041.2 5277874.2, 455041.2 5277866...	0.987419	1
435	POLYGON ((455041.2 5277842.4, 455040.6 5277841...	0.987318	1
469	POLYGON ((455072.4 5277718.2, 455071.8 5277717...	0.986584	1
436	POLYGON ((455003.4 5277738.6, 455002.8 5277738...	0.986499	1

In [12]:

geoai.view_vector_interactive(
    gdf, column="confidence", layer_name="Building", tiles="Satellite"
)

geoai.view_vector_interactive( gdf, column="confidence", layer_name="Building", tiles="Satellite" )

Out[12]:

Make this Notebook Trusted to load map: File -> Trust Notebook

In [13]:

geoai.view_vector_interactive(
    gdf, column="confidence", layer_name="Building", tiles=raster_url
)

geoai.view_vector_interactive( gdf, column="confidence", layer_name="Building", tiles=raster_url )

Out[13]:

Make this Notebook Trusted to load map: File -> Trust Notebook

In [14]:

geoai.view_vector_interactive(
    gdf_regularized, column="confidence", layer_name="Building", tiles=raster_url
)

geoai.view_vector_interactive( gdf_regularized, column="confidence", layer_name="Building", tiles=raster_url )

Out[14]:

Make this Notebook Trusted to load map: File -> Trust Notebook

In [15]:

extractor.visualize_results(raster_path, gdf, output_path="naip_buildings.png")

extractor.visualize_results(raster_path, gdf, output_path="naip_buildings.png")

Using confidence values (range: 0.51 - 0.99)

Visualization saved to naip_buildings.png

Sample visualization saved to naip_buildings_sample.png

In [16]:

extractor.visualize_results(
    raster_path, gdf_regularized, output_path="naip_buildings_regularized.png"
)

extractor.visualize_results( raster_path, gdf_regularized, output_path="naip_buildings_regularized.png" )

Using confidence values (range: 0.51 - 0.99)

Visualization saved to naip_buildings_regularized.png

Sample visualization saved to naip_buildings_regularized_sample.png